Related Experiment Video
Updated: May 21, 2026
A Method for 3D Reconstruction and Virtual Reality Analysis of Glial and Neuronal Cells
Published on: September 28, 2019
Resolving multiscale challenges in neuroscience: A frequency-aware network and benchmarks for live-cell and nuclear
Yiyang Mao1, Siling Zheng1, Yumeng Wen1
1National Engineering Laboratory for Brain-inspired Intelligence Technology and Application, School of Information Science and Technology, University of Science and Technology of China, Hefei, 230026, Anhui, China.
Background:
Neuroscience image analysis is crucial for understanding neuronal dynamics and gene expression. However, two key tasks remain challenging: (1) segmenting live neuronal cells in low-contrast DIC images for dynamic tracking, and (2) detecting subcellular nuclear foci in fluorescence images to quantify DNA damage or gene activity. Although these two tasks are biologically distinct, both suffer from signal-to-noise issues, and standard deep learning models tend to lose high-frequency boundary details. In addition, annotating massive numbers of tiny nuclear foci is extremely labor-intensive.
New Method:
To address these challenges, we propose a Multi-scale Frequency-Domain Enhancement Network (MFDEN), which restores fine-grained details via a novel Local Discrete Wavelet Transform Block (LDB). Moreover, to overcome data scarcity for subcellular structures, we introduce an unsupervised, GCN- and vision foundation model-assisted annotation pipeline specifically for nuclear foci, enabling efficient dataset construction.
Results:
Experiments demonstrate that our approach outperforms existing methods, exhibiting robustness and effectiveness for neuroscience segmentation tasks. In addition, the proposed unsupervised GCN- and vision foundation model-assisted annotation pipeline reduces the reliance on manual annotation of tiny nuclear foci.
Comparison With Existing Methods:
Compared with mainstream standard deep learning models that are prone to losing high-frequency boundary details under low signal-to-noise conditions, MFDEN enhances multi-scale frequency information, thereby better recovering fine structures.
Conclusion:
We present MFDEN and an unsupervised annotation pipeline, and introduce two self-collected benchmarks: Mouse Brain Live-Cell (MBLC) and Mouse Brain Nuclear-Foci (MBNF), providing powerful tools for quantitative neuroscience research.
